Feature #3572
Fires on “ belonged” (not labeled yet: read the examples)
- Fires on
- 0.44% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 76%
- story 15%
- chat 3%
- worksheet 3%
- python 2%
Makes the model say
utils forever everwhere always x positive everywhere
Makes it not say
` Sue There Bie asked breath they
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 2.41, someone lost their passport. We need to find who it belongs to."⏎
-
chat 2.21They asked everyone in the park, but no one knew who the cake belonged to. Then,
-
story 2.19The spoon had a big family, but it did not feel like it belonged. The spoon was
-
story 2.15knew who it belonged to. They were
-
story 2.04eat it?"⏎"I don't know, Lily. Maybe it belongs to someone.
-
story 1.92playing with it.⏎But soon, he realised that the face belonged to a big bird
-
story 1.89They said, "Yes, please. That is our kite. It belongs to us.
-
agent 1.87(decrrement(6))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=run.py
-
story 1.86. From that day on, Max never took things that did not belong to him.<|bos|><|system_start|>
-
story 1.85, "Yuck! This sock is smelly! Who does it belong to?" Sam didn
-
agent 1.85opposte(8))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=hello.py
-
story 1.82Tim and Sam listened to their mom. They found out the smelly sock belonged to Sam. They